Comparison between five classifiers for automatic scoring of human sleep recordings

The aim of this work is to compare the performances of 5 classifiers (linear and quadratic classifiers, k nearest neighbors, Parzen kernels and neural network) to score a set of 8 biological features extracted from EEG and EMG, in six classes corresponding to different sleep stages as to automatically elaborate an hypnogram and help the physician diagnosticate sleep disorders. The data base is composed of 17265 epochs of 20s recorded from 4 patients. Each epoch has been classified by an expert into one of the six sleep stages. In order to evaluate the classifiers, learning and testing sets of fixed size are randomly drawn and are used to train and test the classifiers. After several trials, an estimation of the misclassification percentage and its variability is obtained (optimistically and pessimistically). Data transformations toward normal distribution are explored as an approach to deal with extreme values. It is shown that these transformations improve significantly the results of the classifiers based on data proximity.